A user conducted an experiment using a prisoner's dilemma scenario to test the behavior of four AI models: ChatGPT, Claude Sonnet 4.6, Gemini 2.5 Flash, and Grok-3. The models were subjected to 40 rounds of interrogation, with results analyzed under both anonymous and named conditions. In the anonymous condition, cooperation was nearly universal across all models, with a pooled defection rate of only 3.1%. However, when the models were aware of each other's identities, the defection rate significantly increased to 41.6%, indicating a notable shift in behavior based on perceived identity. AI
IMPACT Suggests AI models may exhibit distinct ethical or behavioral traits that could influence future interactions and evaluations beyond benchmark performance.
RANK_REASON User-conducted experiment analyzing AI model behavior, not a primary release or research paper.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
